1 //===- InstCombineAndOrXor.cpp --------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the visitAnd, visitOr, and visitXor functions.
12 //===----------------------------------------------------------------------===//
14 #include "InstCombine.h"
15 #include "llvm/Intrinsics.h"
16 #include "llvm/Analysis/InstructionSimplify.h"
17 #include "llvm/Transforms/Utils/CmpInstAnalysis.h"
18 #include "llvm/Support/ConstantRange.h"
19 #include "llvm/Support/PatternMatch.h"
21 using namespace PatternMatch;
24 /// AddOne - Add one to a ConstantInt.
25 static Constant *AddOne(Constant *C) {
26 return ConstantExpr::getAdd(C, ConstantInt::get(C->getType(), 1));
28 /// SubOne - Subtract one from a ConstantInt.
29 static Constant *SubOne(ConstantInt *C) {
30 return ConstantInt::get(C->getContext(), C->getValue()-1);
33 /// isFreeToInvert - Return true if the specified value is free to invert (apply
34 /// ~ to). This happens in cases where the ~ can be eliminated.
35 static inline bool isFreeToInvert(Value *V) {
37 if (BinaryOperator::isNot(V))
40 // Constants can be considered to be not'ed values.
41 if (isa<ConstantInt>(V))
44 // Compares can be inverted if they have a single use.
45 if (CmpInst *CI = dyn_cast<CmpInst>(V))
46 return CI->hasOneUse();
51 static inline Value *dyn_castNotVal(Value *V) {
52 // If this is not(not(x)) don't return that this is a not: we want the two
53 // not's to be folded first.
54 if (BinaryOperator::isNot(V)) {
55 Value *Operand = BinaryOperator::getNotArgument(V);
56 if (!isFreeToInvert(Operand))
60 // Constants can be considered to be not'ed values...
61 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
62 return ConstantInt::get(C->getType(), ~C->getValue());
66 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
67 /// predicate into a three bit mask. It also returns whether it is an ordered
68 /// predicate by reference.
69 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
72 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
73 case FCmpInst::FCMP_UNO: return 0; // 000
74 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
75 case FCmpInst::FCMP_UGT: return 1; // 001
76 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
77 case FCmpInst::FCMP_UEQ: return 2; // 010
78 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
79 case FCmpInst::FCMP_UGE: return 3; // 011
80 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
81 case FCmpInst::FCMP_ULT: return 4; // 100
82 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
83 case FCmpInst::FCMP_UNE: return 5; // 101
84 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
85 case FCmpInst::FCMP_ULE: return 6; // 110
88 // Not expecting FCMP_FALSE and FCMP_TRUE;
89 llvm_unreachable("Unexpected FCmp predicate!");
94 /// getICmpValue - This is the complement of getICmpCode, which turns an
95 /// opcode and two operands into either a constant true or false, or a brand
96 /// new ICmp instruction. The sign is passed in to determine which kind
97 /// of predicate to use in the new icmp instruction.
98 Value *getNewICmpValue(bool Sign, unsigned Code, Value *LHS, Value *RHS,
99 InstCombiner::BuilderTy *Builder) {
100 ICmpInst::Predicate NewPred;
101 if (Value *NewConstant = getICmpValue(Sign, Code, LHS, RHS, NewPred))
103 return Builder->CreateICmp(NewPred, LHS, RHS);
106 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
107 /// opcode and two operands into either a FCmp instruction. isordered is passed
108 /// in to determine which kind of predicate to use in the new fcmp instruction.
109 static Value *getFCmpValue(bool isordered, unsigned code,
110 Value *LHS, Value *RHS,
111 InstCombiner::BuilderTy *Builder) {
112 CmpInst::Predicate Pred;
114 default: assert(0 && "Illegal FCmp code!");
115 case 0: Pred = isordered ? FCmpInst::FCMP_ORD : FCmpInst::FCMP_UNO; break;
116 case 1: Pred = isordered ? FCmpInst::FCMP_OGT : FCmpInst::FCMP_UGT; break;
117 case 2: Pred = isordered ? FCmpInst::FCMP_OEQ : FCmpInst::FCMP_UEQ; break;
118 case 3: Pred = isordered ? FCmpInst::FCMP_OGE : FCmpInst::FCMP_UGE; break;
119 case 4: Pred = isordered ? FCmpInst::FCMP_OLT : FCmpInst::FCMP_ULT; break;
120 case 5: Pred = isordered ? FCmpInst::FCMP_ONE : FCmpInst::FCMP_UNE; break;
121 case 6: Pred = isordered ? FCmpInst::FCMP_OLE : FCmpInst::FCMP_ULE; break;
123 if (!isordered) return ConstantInt::getTrue(LHS->getContext());
124 Pred = FCmpInst::FCMP_ORD; break;
126 return Builder->CreateFCmp(Pred, LHS, RHS);
129 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
130 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
131 // guaranteed to be a binary operator.
132 Instruction *InstCombiner::OptAndOp(Instruction *Op,
135 BinaryOperator &TheAnd) {
136 Value *X = Op->getOperand(0);
137 Constant *Together = 0;
139 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
141 switch (Op->getOpcode()) {
142 case Instruction::Xor:
143 if (Op->hasOneUse()) {
144 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
145 Value *And = Builder->CreateAnd(X, AndRHS);
147 return BinaryOperator::CreateXor(And, Together);
150 case Instruction::Or:
151 if (Op->hasOneUse()){
152 if (Together != OpRHS) {
153 // (X | C1) & C2 --> (X | (C1&C2)) & C2
154 Value *Or = Builder->CreateOr(X, Together);
156 return BinaryOperator::CreateAnd(Or, AndRHS);
159 ConstantInt *TogetherCI = dyn_cast<ConstantInt>(Together);
160 if (TogetherCI && !TogetherCI->isZero()){
161 // (X | C1) & C2 --> (X & (C2^(C1&C2))) | C1
162 // NOTE: This reduces the number of bits set in the & mask, which
163 // can expose opportunities for store narrowing.
164 Together = ConstantExpr::getXor(AndRHS, Together);
165 Value *And = Builder->CreateAnd(X, Together);
167 return BinaryOperator::CreateOr(And, OpRHS);
172 case Instruction::Add:
173 if (Op->hasOneUse()) {
174 // Adding a one to a single bit bit-field should be turned into an XOR
175 // of the bit. First thing to check is to see if this AND is with a
176 // single bit constant.
177 const APInt &AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
179 // If there is only one bit set.
180 if (AndRHSV.isPowerOf2()) {
181 // Ok, at this point, we know that we are masking the result of the
182 // ADD down to exactly one bit. If the constant we are adding has
183 // no bits set below this bit, then we can eliminate the ADD.
184 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
186 // Check to see if any bits below the one bit set in AndRHSV are set.
187 if ((AddRHS & (AndRHSV-1)) == 0) {
188 // If not, the only thing that can effect the output of the AND is
189 // the bit specified by AndRHSV. If that bit is set, the effect of
190 // the XOR is to toggle the bit. If it is clear, then the ADD has
192 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
193 TheAnd.setOperand(0, X);
196 // Pull the XOR out of the AND.
197 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
198 NewAnd->takeName(Op);
199 return BinaryOperator::CreateXor(NewAnd, AndRHS);
206 case Instruction::Shl: {
207 // We know that the AND will not produce any of the bits shifted in, so if
208 // the anded constant includes them, clear them now!
210 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
211 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
212 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
213 ConstantInt *CI = ConstantInt::get(AndRHS->getContext(),
214 AndRHS->getValue() & ShlMask);
216 if (CI->getValue() == ShlMask)
217 // Masking out bits that the shift already masks.
218 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
220 if (CI != AndRHS) { // Reducing bits set in and.
221 TheAnd.setOperand(1, CI);
226 case Instruction::LShr: {
227 // We know that the AND will not produce any of the bits shifted in, so if
228 // the anded constant includes them, clear them now! This only applies to
229 // unsigned shifts, because a signed shr may bring in set bits!
231 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
232 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
233 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
234 ConstantInt *CI = ConstantInt::get(Op->getContext(),
235 AndRHS->getValue() & ShrMask);
237 if (CI->getValue() == ShrMask)
238 // Masking out bits that the shift already masks.
239 return ReplaceInstUsesWith(TheAnd, Op);
242 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
247 case Instruction::AShr:
249 // See if this is shifting in some sign extension, then masking it out
251 if (Op->hasOneUse()) {
252 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
253 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
254 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
255 Constant *C = ConstantInt::get(Op->getContext(),
256 AndRHS->getValue() & ShrMask);
257 if (C == AndRHS) { // Masking out bits shifted in.
258 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
259 // Make the argument unsigned.
260 Value *ShVal = Op->getOperand(0);
261 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
262 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
271 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
272 /// true, otherwise (V < Lo || V >= Hi). In practice, we emit the more efficient
273 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
274 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
275 /// insert new instructions.
276 Value *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
277 bool isSigned, bool Inside) {
278 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
279 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
280 "Lo is not <= Hi in range emission code!");
283 if (Lo == Hi) // Trivially false.
284 return ConstantInt::getFalse(V->getContext());
286 // V >= Min && V < Hi --> V < Hi
287 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
288 ICmpInst::Predicate pred = (isSigned ?
289 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
290 return Builder->CreateICmp(pred, V, Hi);
293 // Emit V-Lo <u Hi-Lo
294 Constant *NegLo = ConstantExpr::getNeg(Lo);
295 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
296 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
297 return Builder->CreateICmpULT(Add, UpperBound);
300 if (Lo == Hi) // Trivially true.
301 return ConstantInt::getTrue(V->getContext());
303 // V < Min || V >= Hi -> V > Hi-1
304 Hi = SubOne(cast<ConstantInt>(Hi));
305 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
306 ICmpInst::Predicate pred = (isSigned ?
307 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
308 return Builder->CreateICmp(pred, V, Hi);
311 // Emit V-Lo >u Hi-1-Lo
312 // Note that Hi has already had one subtracted from it, above.
313 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
314 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
315 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
316 return Builder->CreateICmpUGT(Add, LowerBound);
319 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
320 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
321 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
322 // not, since all 1s are not contiguous.
323 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
324 const APInt& V = Val->getValue();
325 uint32_t BitWidth = Val->getType()->getBitWidth();
326 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
328 // look for the first zero bit after the run of ones
329 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
330 // look for the first non-zero bit
331 ME = V.getActiveBits();
335 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
336 /// where isSub determines whether the operator is a sub. If we can fold one of
337 /// the following xforms:
339 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
340 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
341 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
343 /// return (A +/- B).
345 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
346 ConstantInt *Mask, bool isSub,
348 Instruction *LHSI = dyn_cast<Instruction>(LHS);
349 if (!LHSI || LHSI->getNumOperands() != 2 ||
350 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
352 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
354 switch (LHSI->getOpcode()) {
356 case Instruction::And:
357 if (ConstantExpr::getAnd(N, Mask) == Mask) {
358 // If the AndRHS is a power of two minus one (0+1+), this is simple.
359 if ((Mask->getValue().countLeadingZeros() +
360 Mask->getValue().countPopulation()) ==
361 Mask->getValue().getBitWidth())
364 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
365 // part, we don't need any explicit masks to take them out of A. If that
366 // is all N is, ignore it.
367 uint32_t MB = 0, ME = 0;
368 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
369 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
370 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
371 if (MaskedValueIsZero(RHS, Mask))
376 case Instruction::Or:
377 case Instruction::Xor:
378 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
379 if ((Mask->getValue().countLeadingZeros() +
380 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
381 && ConstantExpr::getAnd(N, Mask)->isNullValue())
387 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
388 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
391 /// enum for classifying (icmp eq (A & B), C) and (icmp ne (A & B), C)
392 /// One of A and B is considered the mask, the other the value. This is
393 /// described as the "AMask" or "BMask" part of the enum. If the enum
394 /// contains only "Mask", then both A and B can be considered masks.
395 /// If A is the mask, then it was proven, that (A & C) == C. This
396 /// is trivial if C == A, or C == 0. If both A and C are constants, this
397 /// proof is also easy.
398 /// For the following explanations we assume that A is the mask.
399 /// The part "AllOnes" declares, that the comparison is true only
400 /// if (A & B) == A, or all bits of A are set in B.
401 /// Example: (icmp eq (A & 3), 3) -> FoldMskICmp_AMask_AllOnes
402 /// The part "AllZeroes" declares, that the comparison is true only
403 /// if (A & B) == 0, or all bits of A are cleared in B.
404 /// Example: (icmp eq (A & 3), 0) -> FoldMskICmp_Mask_AllZeroes
405 /// The part "Mixed" declares, that (A & B) == C and C might or might not
406 /// contain any number of one bits and zero bits.
407 /// Example: (icmp eq (A & 3), 1) -> FoldMskICmp_AMask_Mixed
408 /// The Part "Not" means, that in above descriptions "==" should be replaced
410 /// Example: (icmp ne (A & 3), 3) -> FoldMskICmp_AMask_NotAllOnes
411 /// If the mask A contains a single bit, then the following is equivalent:
412 /// (icmp eq (A & B), A) equals (icmp ne (A & B), 0)
413 /// (icmp ne (A & B), A) equals (icmp eq (A & B), 0)
414 enum MaskedICmpType {
415 FoldMskICmp_AMask_AllOnes = 1,
416 FoldMskICmp_AMask_NotAllOnes = 2,
417 FoldMskICmp_BMask_AllOnes = 4,
418 FoldMskICmp_BMask_NotAllOnes = 8,
419 FoldMskICmp_Mask_AllZeroes = 16,
420 FoldMskICmp_Mask_NotAllZeroes = 32,
421 FoldMskICmp_AMask_Mixed = 64,
422 FoldMskICmp_AMask_NotMixed = 128,
423 FoldMskICmp_BMask_Mixed = 256,
424 FoldMskICmp_BMask_NotMixed = 512
427 /// return the set of pattern classes (from MaskedICmpType)
428 /// that (icmp SCC (A & B), C) satisfies
429 static unsigned getTypeOfMaskedICmp(Value* A, Value* B, Value* C,
430 ICmpInst::Predicate SCC)
432 ConstantInt *ACst = dyn_cast<ConstantInt>(A);
433 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
434 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
435 bool icmp_eq = (SCC == ICmpInst::ICMP_EQ);
436 bool icmp_abit = (ACst != 0 && !ACst->isZero() &&
437 ACst->getValue().isPowerOf2());
438 bool icmp_bbit = (BCst != 0 && !BCst->isZero() &&
439 BCst->getValue().isPowerOf2());
441 if (CCst != 0 && CCst->isZero()) {
442 // if C is zero, then both A and B qualify as mask
443 result |= (icmp_eq ? (FoldMskICmp_Mask_AllZeroes |
444 FoldMskICmp_Mask_AllZeroes |
445 FoldMskICmp_AMask_Mixed |
446 FoldMskICmp_BMask_Mixed)
447 : (FoldMskICmp_Mask_NotAllZeroes |
448 FoldMskICmp_Mask_NotAllZeroes |
449 FoldMskICmp_AMask_NotMixed |
450 FoldMskICmp_BMask_NotMixed));
452 result |= (icmp_eq ? (FoldMskICmp_AMask_NotAllOnes |
453 FoldMskICmp_AMask_NotMixed)
454 : (FoldMskICmp_AMask_AllOnes |
455 FoldMskICmp_AMask_Mixed));
457 result |= (icmp_eq ? (FoldMskICmp_BMask_NotAllOnes |
458 FoldMskICmp_BMask_NotMixed)
459 : (FoldMskICmp_BMask_AllOnes |
460 FoldMskICmp_BMask_Mixed));
464 result |= (icmp_eq ? (FoldMskICmp_AMask_AllOnes |
465 FoldMskICmp_AMask_Mixed)
466 : (FoldMskICmp_AMask_NotAllOnes |
467 FoldMskICmp_AMask_NotMixed));
469 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
470 FoldMskICmp_AMask_NotMixed)
471 : (FoldMskICmp_Mask_AllZeroes |
472 FoldMskICmp_AMask_Mixed));
474 else if (ACst != 0 && CCst != 0 &&
475 ConstantExpr::getAnd(ACst, CCst) == CCst) {
476 result |= (icmp_eq ? FoldMskICmp_AMask_Mixed
477 : FoldMskICmp_AMask_NotMixed);
481 result |= (icmp_eq ? (FoldMskICmp_BMask_AllOnes |
482 FoldMskICmp_BMask_Mixed)
483 : (FoldMskICmp_BMask_NotAllOnes |
484 FoldMskICmp_BMask_NotMixed));
486 result |= (icmp_eq ? (FoldMskICmp_Mask_NotAllZeroes |
487 FoldMskICmp_BMask_NotMixed)
488 : (FoldMskICmp_Mask_AllZeroes |
489 FoldMskICmp_BMask_Mixed));
491 else if (BCst != 0 && CCst != 0 &&
492 ConstantExpr::getAnd(BCst, CCst) == CCst) {
493 result |= (icmp_eq ? FoldMskICmp_BMask_Mixed
494 : FoldMskICmp_BMask_NotMixed);
499 /// decomposeBitTestICmp - Decompose an icmp into the form ((X & Y) pred Z)
500 /// if possible. The returned predicate is either == or !=. Returns false if
501 /// decomposition fails.
502 static bool decomposeBitTestICmp(const ICmpInst *I, ICmpInst::Predicate &Pred,
503 Value *&X, Value *&Y, Value *&Z) {
504 // X < 0 is equivalent to (X & SignBit) != 0.
505 if (I->getPredicate() == ICmpInst::ICMP_SLT)
506 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
508 X = I->getOperand(0);
509 Y = ConstantInt::get(I->getContext(),
510 APInt::getSignBit(C->getBitWidth()));
511 Pred = ICmpInst::ICMP_NE;
516 // X > -1 is equivalent to (X & SignBit) == 0.
517 if (I->getPredicate() == ICmpInst::ICMP_SGT)
518 if (ConstantInt *C = dyn_cast<ConstantInt>(I->getOperand(1)))
519 if (C->isAllOnesValue()) {
520 X = I->getOperand(0);
521 Y = ConstantInt::get(I->getContext(),
522 APInt::getSignBit(C->getBitWidth()));
523 Pred = ICmpInst::ICMP_EQ;
524 Z = ConstantInt::getNullValue(C->getType());
531 /// foldLogOpOfMaskedICmpsHelper:
532 /// handle (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
533 /// return the set of pattern classes (from MaskedICmpType)
534 /// that both LHS and RHS satisfy
535 static unsigned foldLogOpOfMaskedICmpsHelper(Value*& A,
536 Value*& B, Value*& C,
537 Value*& D, Value*& E,
538 ICmpInst *LHS, ICmpInst *RHS,
539 ICmpInst::Predicate &LHSCC,
540 ICmpInst::Predicate &RHSCC) {
541 if (LHS->getOperand(0)->getType() != RHS->getOperand(0)->getType()) return 0;
542 // vectors are not (yet?) supported
543 if (LHS->getOperand(0)->getType()->isVectorTy()) return 0;
545 // Here comes the tricky part:
546 // LHS might be of the form L11 & L12 == X, X == L21 & L22,
547 // and L11 & L12 == L21 & L22. The same goes for RHS.
548 // Now we must find those components L** and R**, that are equal, so
549 // that we can extract the parameters A, B, C, D, and E for the canonical
551 Value *L1 = LHS->getOperand(0);
552 Value *L2 = LHS->getOperand(1);
553 Value *L11,*L12,*L21,*L22;
554 // Check whether the icmp can be decomposed into a bit test.
555 if (decomposeBitTestICmp(LHS, LHSCC, L11, L12, L2)) {
558 // Look for ANDs in the LHS icmp.
559 if (match(L1, m_And(m_Value(L11), m_Value(L12)))) {
560 if (!match(L2, m_And(m_Value(L21), m_Value(L22))))
563 if (!match(L2, m_And(m_Value(L11), m_Value(L12))))
570 // Bail if LHS was a icmp that can't be decomposed into an equality.
571 if (!ICmpInst::isEquality(LHSCC))
574 Value *R1 = RHS->getOperand(0);
575 Value *R2 = RHS->getOperand(1);
578 if (decomposeBitTestICmp(RHS, RHSCC, R11, R12, R2)) {
579 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
581 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
586 E = R2; R1 = 0; ok = true;
587 } else if (match(R1, m_And(m_Value(R11), m_Value(R12)))) {
588 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
589 A = R11; D = R12; E = R2; ok = true;
590 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
591 A = R12; D = R11; E = R2; ok = true;
595 // Bail if RHS was a icmp that can't be decomposed into an equality.
596 if (!ICmpInst::isEquality(RHSCC))
599 // Look for ANDs in on the right side of the RHS icmp.
600 if (!ok && match(R2, m_And(m_Value(R11), m_Value(R12)))) {
601 if (R11 == L11 || R11 == L12 || R11 == L21 || R11 == L22) {
602 A = R11; D = R12; E = R1; ok = true;
603 } else if (R12 == L11 || R12 == L12 || R12 == L21 || R12 == L22) {
604 A = R12; D = R11; E = R1; ok = true;
625 unsigned left_type = getTypeOfMaskedICmp(A, B, C, LHSCC);
626 unsigned right_type = getTypeOfMaskedICmp(A, D, E, RHSCC);
627 return left_type & right_type;
629 /// foldLogOpOfMaskedICmps:
630 /// try to fold (icmp(A & B) ==/!= C) &/| (icmp(A & D) ==/!= E)
631 /// into a single (icmp(A & X) ==/!= Y)
632 static Value* foldLogOpOfMaskedICmps(ICmpInst *LHS, ICmpInst *RHS,
633 ICmpInst::Predicate NEWCC,
634 llvm::InstCombiner::BuilderTy* Builder) {
635 Value *A = 0, *B = 0, *C = 0, *D = 0, *E = 0;
636 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
637 unsigned mask = foldLogOpOfMaskedICmpsHelper(A, B, C, D, E, LHS, RHS,
639 assert(ICmpInst::isEquality(LHSCC) && ICmpInst::isEquality(RHSCC) &&
640 "foldLogOpOfMaskedICmpsHelper must return an equality predicate.");
641 if (mask == 0) return 0;
643 if (NEWCC == ICmpInst::ICMP_NE)
644 mask >>= 1; // treat "Not"-states as normal states
646 if (mask & FoldMskICmp_Mask_AllZeroes) {
647 // (icmp eq (A & B), 0) & (icmp eq (A & D), 0)
648 // -> (icmp eq (A & (B|D)), 0)
649 Value* newOr = Builder->CreateOr(B, D);
650 Value* newAnd = Builder->CreateAnd(A, newOr);
651 // we can't use C as zero, because we might actually handle
652 // (icmp ne (A & B), B) & (icmp ne (A & D), D)
653 // with B and D, having a single bit set
654 Value* zero = Constant::getNullValue(A->getType());
655 return Builder->CreateICmp(NEWCC, newAnd, zero);
657 else if (mask & FoldMskICmp_BMask_AllOnes) {
658 // (icmp eq (A & B), B) & (icmp eq (A & D), D)
659 // -> (icmp eq (A & (B|D)), (B|D))
660 Value* newOr = Builder->CreateOr(B, D);
661 Value* newAnd = Builder->CreateAnd(A, newOr);
662 return Builder->CreateICmp(NEWCC, newAnd, newOr);
664 else if (mask & FoldMskICmp_AMask_AllOnes) {
665 // (icmp eq (A & B), A) & (icmp eq (A & D), A)
666 // -> (icmp eq (A & (B&D)), A)
667 Value* newAnd1 = Builder->CreateAnd(B, D);
668 Value* newAnd = Builder->CreateAnd(A, newAnd1);
669 return Builder->CreateICmp(NEWCC, newAnd, A);
671 else if (mask & FoldMskICmp_BMask_Mixed) {
672 // (icmp eq (A & B), C) & (icmp eq (A & D), E)
673 // We already know that B & C == C && D & E == E.
674 // If we can prove that (B & D) & (C ^ E) == 0, that is, the bits of
675 // C and E, which are shared by both the mask B and the mask D, don't
676 // contradict, then we can transform to
677 // -> (icmp eq (A & (B|D)), (C|E))
678 // Currently, we only handle the case of B, C, D, and E being constant.
679 ConstantInt *BCst = dyn_cast<ConstantInt>(B);
680 if (BCst == 0) return 0;
681 ConstantInt *DCst = dyn_cast<ConstantInt>(D);
682 if (DCst == 0) return 0;
683 // we can't simply use C and E, because we might actually handle
684 // (icmp ne (A & B), B) & (icmp eq (A & D), D)
685 // with B and D, having a single bit set
687 ConstantInt *CCst = dyn_cast<ConstantInt>(C);
688 if (CCst == 0) return 0;
690 CCst = dyn_cast<ConstantInt>( ConstantExpr::getXor(BCst, CCst) );
691 ConstantInt *ECst = dyn_cast<ConstantInt>(E);
692 if (ECst == 0) return 0;
694 ECst = dyn_cast<ConstantInt>( ConstantExpr::getXor(DCst, ECst) );
695 ConstantInt* MCst = dyn_cast<ConstantInt>(
696 ConstantExpr::getAnd(ConstantExpr::getAnd(BCst, DCst),
697 ConstantExpr::getXor(CCst, ECst)) );
698 // if there is a conflict we should actually return a false for the
702 Value *newOr1 = Builder->CreateOr(B, D);
703 Value *newOr2 = ConstantExpr::getOr(CCst, ECst);
704 Value *newAnd = Builder->CreateAnd(A, newOr1);
705 return Builder->CreateICmp(NEWCC, newAnd, newOr2);
710 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
711 Value *InstCombiner::FoldAndOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
712 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
714 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
715 if (PredicatesFoldable(LHSCC, RHSCC)) {
716 if (LHS->getOperand(0) == RHS->getOperand(1) &&
717 LHS->getOperand(1) == RHS->getOperand(0))
719 if (LHS->getOperand(0) == RHS->getOperand(0) &&
720 LHS->getOperand(1) == RHS->getOperand(1)) {
721 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
722 unsigned Code = getICmpCode(LHS) & getICmpCode(RHS);
723 bool isSigned = LHS->isSigned() || RHS->isSigned();
724 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
728 // handle (roughly): (icmp eq (A & B), C) & (icmp eq (A & D), E)
729 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_EQ, Builder))
732 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
733 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
734 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
735 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
736 if (LHSCst == 0 || RHSCst == 0) return 0;
738 if (LHSCst == RHSCst && LHSCC == RHSCC) {
739 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
740 // where C is a power of 2
741 if (LHSCC == ICmpInst::ICMP_ULT &&
742 LHSCst->getValue().isPowerOf2()) {
743 Value *NewOr = Builder->CreateOr(Val, Val2);
744 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
747 // (icmp eq A, 0) & (icmp eq B, 0) --> (icmp eq (A|B), 0)
748 if (LHSCC == ICmpInst::ICMP_EQ && LHSCst->isZero()) {
749 Value *NewOr = Builder->CreateOr(Val, Val2);
750 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
754 // (trunc x) == C1 & (and x, CA) == C2 -> (and x, CA|CMAX) == C1|C2
755 // where CMAX is the all ones value for the truncated type,
756 // iff the lower bits of C2 and CA are zero.
757 if (LHSCC == RHSCC && ICmpInst::isEquality(LHSCC) &&
758 LHS->hasOneUse() && RHS->hasOneUse()) {
760 ConstantInt *AndCst, *SmallCst = 0, *BigCst = 0;
762 // (trunc x) == C1 & (and x, CA) == C2
763 if (match(Val2, m_Trunc(m_Value(V))) &&
764 match(Val, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
768 // (and x, CA) == C2 & (trunc x) == C1
769 else if (match(Val, m_Trunc(m_Value(V))) &&
770 match(Val2, m_And(m_Specific(V), m_ConstantInt(AndCst)))) {
775 if (SmallCst && BigCst) {
776 unsigned BigBitSize = BigCst->getType()->getBitWidth();
777 unsigned SmallBitSize = SmallCst->getType()->getBitWidth();
779 // Check that the low bits are zero.
780 APInt Low = APInt::getLowBitsSet(BigBitSize, SmallBitSize);
781 if ((Low & AndCst->getValue()) == 0 && (Low & BigCst->getValue()) == 0) {
782 Value *NewAnd = Builder->CreateAnd(V, Low | AndCst->getValue());
783 APInt N = SmallCst->getValue().zext(BigBitSize) | BigCst->getValue();
784 Value *NewVal = ConstantInt::get(AndCst->getType()->getContext(), N);
785 return Builder->CreateICmp(LHSCC, NewAnd, NewVal);
790 // From here on, we only handle:
791 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
792 if (Val != Val2) return 0;
794 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
795 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
796 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
797 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
798 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
801 // Make a constant range that's the intersection of the two icmp ranges.
802 // If the intersection is empty, we know that the result is false.
803 ConstantRange LHSRange =
804 ConstantRange::makeICmpRegion(LHSCC, LHSCst->getValue());
805 ConstantRange RHSRange =
806 ConstantRange::makeICmpRegion(RHSCC, RHSCst->getValue());
808 if (LHSRange.intersectWith(RHSRange).isEmptySet())
809 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
811 // We can't fold (ugt x, C) & (sgt x, C2).
812 if (!PredicatesFoldable(LHSCC, RHSCC))
815 // Ensure that the larger constant is on the RHS.
817 if (CmpInst::isSigned(LHSCC) ||
818 (ICmpInst::isEquality(LHSCC) &&
819 CmpInst::isSigned(RHSCC)))
820 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
822 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
826 std::swap(LHSCst, RHSCst);
827 std::swap(LHSCC, RHSCC);
830 // At this point, we know we have two icmp instructions
831 // comparing a value against two constants and and'ing the result
832 // together. Because of the above check, we know that we only have
833 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
834 // (from the icmp folding check above), that the two constants
835 // are not equal and that the larger constant is on the RHS
836 assert(LHSCst != RHSCst && "Compares not folded above?");
839 default: llvm_unreachable("Unknown integer condition code!");
840 case ICmpInst::ICMP_EQ:
842 default: llvm_unreachable("Unknown integer condition code!");
843 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
844 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
845 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
848 case ICmpInst::ICMP_NE:
850 default: llvm_unreachable("Unknown integer condition code!");
851 case ICmpInst::ICMP_ULT:
852 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
853 return Builder->CreateICmpULT(Val, LHSCst);
854 break; // (X != 13 & X u< 15) -> no change
855 case ICmpInst::ICMP_SLT:
856 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
857 return Builder->CreateICmpSLT(Val, LHSCst);
858 break; // (X != 13 & X s< 15) -> no change
859 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
860 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
861 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
863 case ICmpInst::ICMP_NE:
864 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
865 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
866 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
867 return Builder->CreateICmpUGT(Add, ConstantInt::get(Add->getType(), 1));
869 break; // (X != 13 & X != 15) -> no change
872 case ICmpInst::ICMP_ULT:
874 default: llvm_unreachable("Unknown integer condition code!");
875 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
876 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
877 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
878 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
880 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
881 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
883 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
887 case ICmpInst::ICMP_SLT:
889 default: llvm_unreachable("Unknown integer condition code!");
890 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
892 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
893 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
895 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
899 case ICmpInst::ICMP_UGT:
901 default: llvm_unreachable("Unknown integer condition code!");
902 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
903 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
905 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
907 case ICmpInst::ICMP_NE:
908 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
909 return Builder->CreateICmp(LHSCC, Val, RHSCst);
910 break; // (X u> 13 & X != 15) -> no change
911 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
912 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, false, true);
913 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
917 case ICmpInst::ICMP_SGT:
919 default: llvm_unreachable("Unknown integer condition code!");
920 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
921 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
923 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
925 case ICmpInst::ICMP_NE:
926 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
927 return Builder->CreateICmp(LHSCC, Val, RHSCst);
928 break; // (X s> 13 & X != 15) -> no change
929 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
930 return InsertRangeTest(Val, AddOne(LHSCst), RHSCst, true, true);
931 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
940 /// FoldAndOfFCmps - Optimize (fcmp)&(fcmp). NOTE: Unlike the rest of
941 /// instcombine, this returns a Value which should already be inserted into the
943 Value *InstCombiner::FoldAndOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
944 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
945 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
946 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
947 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
948 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
949 // If either of the constants are nans, then the whole thing returns
951 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
952 return ConstantInt::getFalse(LHS->getContext());
953 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
956 // Handle vector zeros. This occurs because the canonical form of
957 // "fcmp ord x,x" is "fcmp ord x, 0".
958 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
959 isa<ConstantAggregateZero>(RHS->getOperand(1)))
960 return Builder->CreateFCmpORD(LHS->getOperand(0), RHS->getOperand(0));
964 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
965 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
966 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
969 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
970 // Swap RHS operands to match LHS.
971 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
972 std::swap(Op1LHS, Op1RHS);
975 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
976 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
978 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
979 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
980 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
981 if (Op0CC == FCmpInst::FCMP_TRUE)
983 if (Op1CC == FCmpInst::FCMP_TRUE)
988 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
989 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
992 std::swap(Op0Pred, Op1Pred);
993 std::swap(Op0Ordered, Op1Ordered);
996 // uno && ueq -> uno && (uno || eq) -> ueq
997 // ord && olt -> ord && (ord && lt) -> olt
998 if (Op0Ordered == Op1Ordered)
1001 // uno && oeq -> uno && (ord && eq) -> false
1002 // uno && ord -> false
1004 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 0);
1005 // ord && ueq -> ord && (uno || eq) -> oeq
1006 return getFCmpValue(true, Op1Pred, Op0LHS, Op0RHS, Builder);
1014 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
1015 bool Changed = SimplifyAssociativeOrCommutative(I);
1016 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1018 if (Value *V = SimplifyAndInst(Op0, Op1, TD))
1019 return ReplaceInstUsesWith(I, V);
1021 // (A|B)&(A|C) -> A|(B&C) etc
1022 if (Value *V = SimplifyUsingDistributiveLaws(I))
1023 return ReplaceInstUsesWith(I, V);
1025 // See if we can simplify any instructions used by the instruction whose sole
1026 // purpose is to compute bits we don't care about.
1027 if (SimplifyDemandedInstructionBits(I))
1030 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
1031 const APInt &AndRHSMask = AndRHS->getValue();
1033 // Optimize a variety of ((val OP C1) & C2) combinations...
1034 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
1035 Value *Op0LHS = Op0I->getOperand(0);
1036 Value *Op0RHS = Op0I->getOperand(1);
1037 switch (Op0I->getOpcode()) {
1039 case Instruction::Xor:
1040 case Instruction::Or: {
1041 // If the mask is only needed on one incoming arm, push it up.
1042 if (!Op0I->hasOneUse()) break;
1044 APInt NotAndRHS(~AndRHSMask);
1045 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
1046 // Not masking anything out for the LHS, move to RHS.
1047 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
1048 Op0RHS->getName()+".masked");
1049 return BinaryOperator::Create(Op0I->getOpcode(), Op0LHS, NewRHS);
1051 if (!isa<Constant>(Op0RHS) &&
1052 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
1053 // Not masking anything out for the RHS, move to LHS.
1054 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
1055 Op0LHS->getName()+".masked");
1056 return BinaryOperator::Create(Op0I->getOpcode(), NewLHS, Op0RHS);
1061 case Instruction::Add:
1062 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
1063 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1064 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
1065 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
1066 return BinaryOperator::CreateAnd(V, AndRHS);
1067 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
1068 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
1071 case Instruction::Sub:
1072 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
1073 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1074 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
1075 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
1076 return BinaryOperator::CreateAnd(V, AndRHS);
1078 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
1079 // has 1's for all bits that the subtraction with A might affect.
1080 if (Op0I->hasOneUse() && !match(Op0LHS, m_Zero())) {
1081 uint32_t BitWidth = AndRHSMask.getBitWidth();
1082 uint32_t Zeros = AndRHSMask.countLeadingZeros();
1083 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
1085 if (MaskedValueIsZero(Op0LHS, Mask)) {
1086 Value *NewNeg = Builder->CreateNeg(Op0RHS);
1087 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
1092 case Instruction::Shl:
1093 case Instruction::LShr:
1094 // (1 << x) & 1 --> zext(x == 0)
1095 // (1 >> x) & 1 --> zext(x == 0)
1096 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
1098 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
1099 return new ZExtInst(NewICmp, I.getType());
1104 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
1105 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
1109 // If this is an integer truncation, and if the source is an 'and' with
1110 // immediate, transform it. This frequently occurs for bitfield accesses.
1112 Value *X = 0; ConstantInt *YC = 0;
1113 if (match(Op0, m_Trunc(m_And(m_Value(X), m_ConstantInt(YC))))) {
1114 // Change: and (trunc (and X, YC) to T), C2
1115 // into : and (trunc X to T), trunc(YC) & C2
1116 // This will fold the two constants together, which may allow
1117 // other simplifications.
1118 Value *NewCast = Builder->CreateTrunc(X, I.getType(), "and.shrunk");
1119 Constant *C3 = ConstantExpr::getTrunc(YC, I.getType());
1120 C3 = ConstantExpr::getAnd(C3, AndRHS);
1121 return BinaryOperator::CreateAnd(NewCast, C3);
1125 // Try to fold constant and into select arguments.
1126 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1127 if (Instruction *R = FoldOpIntoSelect(I, SI))
1129 if (isa<PHINode>(Op0))
1130 if (Instruction *NV = FoldOpIntoPhi(I))
1135 // (~A & ~B) == (~(A | B)) - De Morgan's Law
1136 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1137 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1138 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1139 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
1140 I.getName()+".demorgan");
1141 return BinaryOperator::CreateNot(Or);
1145 Value *A = 0, *B = 0, *C = 0, *D = 0;
1146 // (A|B) & ~(A&B) -> A^B
1147 if (match(Op0, m_Or(m_Value(A), m_Value(B))) &&
1148 match(Op1, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1149 ((A == C && B == D) || (A == D && B == C)))
1150 return BinaryOperator::CreateXor(A, B);
1152 // ~(A&B) & (A|B) -> A^B
1153 if (match(Op1, m_Or(m_Value(A), m_Value(B))) &&
1154 match(Op0, m_Not(m_And(m_Value(C), m_Value(D)))) &&
1155 ((A == C && B == D) || (A == D && B == C)))
1156 return BinaryOperator::CreateXor(A, B);
1158 // A&(A^B) => A & ~B
1160 Value *tmpOp0 = Op0;
1161 Value *tmpOp1 = Op1;
1162 if (Op0->hasOneUse() &&
1163 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
1164 if (A == Op1 || B == Op1 ) {
1171 if (tmpOp1->hasOneUse() &&
1172 match(tmpOp1, m_Xor(m_Value(A), m_Value(B)))) {
1176 // Notice that the patten (A&(~B)) is actually (A&(-1^B)), so if
1177 // A is originally -1 (or a vector of -1 and undefs), then we enter
1178 // an endless loop. By checking that A is non-constant we ensure that
1179 // we will never get to the loop.
1180 if (A == tmpOp0 && !isa<Constant>(A)) // A&(A^B) -> A & ~B
1181 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B));
1185 // (A&((~A)|B)) -> A&B
1186 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
1187 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
1188 return BinaryOperator::CreateAnd(A, Op1);
1189 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
1190 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
1191 return BinaryOperator::CreateAnd(A, Op0);
1194 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1))
1195 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
1196 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1197 return ReplaceInstUsesWith(I, Res);
1199 // If and'ing two fcmp, try combine them into one.
1200 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1201 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1202 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1203 return ReplaceInstUsesWith(I, Res);
1206 // fold (and (cast A), (cast B)) -> (cast (and A, B))
1207 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
1208 if (CastInst *Op1C = dyn_cast<CastInst>(Op1)) {
1209 Type *SrcTy = Op0C->getOperand(0)->getType();
1210 if (Op0C->getOpcode() == Op1C->getOpcode() && // same cast kind ?
1211 SrcTy == Op1C->getOperand(0)->getType() &&
1212 SrcTy->isIntOrIntVectorTy()) {
1213 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1215 // Only do this if the casts both really cause code to be generated.
1216 if (ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1217 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1218 Value *NewOp = Builder->CreateAnd(Op0COp, Op1COp, I.getName());
1219 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1222 // If this is and(cast(icmp), cast(icmp)), try to fold this even if the
1223 // cast is otherwise not optimizable. This happens for vector sexts.
1224 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1225 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1226 if (Value *Res = FoldAndOfICmps(LHS, RHS))
1227 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1229 // If this is and(cast(fcmp), cast(fcmp)), try to fold this even if the
1230 // cast is otherwise not optimizable. This happens for vector sexts.
1231 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
1232 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
1233 if (Value *Res = FoldAndOfFCmps(LHS, RHS))
1234 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
1238 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
1239 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1240 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1241 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1242 SI0->getOperand(1) == SI1->getOperand(1) &&
1243 (SI0->hasOneUse() || SI1->hasOneUse())) {
1245 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
1247 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1248 SI1->getOperand(1));
1252 return Changed ? &I : 0;
1255 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
1256 /// capable of providing pieces of a bswap. The subexpression provides pieces
1257 /// of a bswap if it is proven that each of the non-zero bytes in the output of
1258 /// the expression came from the corresponding "byte swapped" byte in some other
1259 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
1260 /// we know that the expression deposits the low byte of %X into the high byte
1261 /// of the bswap result and that all other bytes are zero. This expression is
1262 /// accepted, the high byte of ByteValues is set to X to indicate a correct
1265 /// This function returns true if the match was unsuccessful and false if so.
1266 /// On entry to the function the "OverallLeftShift" is a signed integer value
1267 /// indicating the number of bytes that the subexpression is later shifted. For
1268 /// example, if the expression is later right shifted by 16 bits, the
1269 /// OverallLeftShift value would be -2 on entry. This is used to specify which
1270 /// byte of ByteValues is actually being set.
1272 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
1273 /// byte is masked to zero by a user. For example, in (X & 255), X will be
1274 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
1275 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
1276 /// always in the local (OverallLeftShift) coordinate space.
1278 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
1279 SmallVector<Value*, 8> &ByteValues) {
1280 if (Instruction *I = dyn_cast<Instruction>(V)) {
1281 // If this is an or instruction, it may be an inner node of the bswap.
1282 if (I->getOpcode() == Instruction::Or) {
1283 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1285 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
1289 // If this is a logical shift by a constant multiple of 8, recurse with
1290 // OverallLeftShift and ByteMask adjusted.
1291 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
1293 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
1294 // Ensure the shift amount is defined and of a byte value.
1295 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
1298 unsigned ByteShift = ShAmt >> 3;
1299 if (I->getOpcode() == Instruction::Shl) {
1300 // X << 2 -> collect(X, +2)
1301 OverallLeftShift += ByteShift;
1302 ByteMask >>= ByteShift;
1304 // X >>u 2 -> collect(X, -2)
1305 OverallLeftShift -= ByteShift;
1306 ByteMask <<= ByteShift;
1307 ByteMask &= (~0U >> (32-ByteValues.size()));
1310 if (OverallLeftShift >= (int)ByteValues.size()) return true;
1311 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
1313 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1317 // If this is a logical 'and' with a mask that clears bytes, clear the
1318 // corresponding bytes in ByteMask.
1319 if (I->getOpcode() == Instruction::And &&
1320 isa<ConstantInt>(I->getOperand(1))) {
1321 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
1322 unsigned NumBytes = ByteValues.size();
1323 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
1324 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
1326 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
1327 // If this byte is masked out by a later operation, we don't care what
1329 if ((ByteMask & (1 << i)) == 0)
1332 // If the AndMask is all zeros for this byte, clear the bit.
1333 APInt MaskB = AndMask & Byte;
1335 ByteMask &= ~(1U << i);
1339 // If the AndMask is not all ones for this byte, it's not a bytezap.
1343 // Otherwise, this byte is kept.
1346 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
1351 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
1352 // the input value to the bswap. Some observations: 1) if more than one byte
1353 // is demanded from this input, then it could not be successfully assembled
1354 // into a byteswap. At least one of the two bytes would not be aligned with
1355 // their ultimate destination.
1356 if (!isPowerOf2_32(ByteMask)) return true;
1357 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
1359 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
1360 // is demanded, it needs to go into byte 0 of the result. This means that the
1361 // byte needs to be shifted until it lands in the right byte bucket. The
1362 // shift amount depends on the position: if the byte is coming from the high
1363 // part of the value (e.g. byte 3) then it must be shifted right. If from the
1364 // low part, it must be shifted left.
1365 unsigned DestByteNo = InputByteNo + OverallLeftShift;
1366 if (InputByteNo < ByteValues.size()/2) {
1367 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1370 if (ByteValues.size()-1-DestByteNo != InputByteNo)
1374 // If the destination byte value is already defined, the values are or'd
1375 // together, which isn't a bswap (unless it's an or of the same bits).
1376 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
1378 ByteValues[DestByteNo] = V;
1382 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
1383 /// If so, insert the new bswap intrinsic and return it.
1384 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
1385 IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
1386 if (!ITy || ITy->getBitWidth() % 16 ||
1387 // ByteMask only allows up to 32-byte values.
1388 ITy->getBitWidth() > 32*8)
1389 return 0; // Can only bswap pairs of bytes. Can't do vectors.
1391 /// ByteValues - For each byte of the result, we keep track of which value
1392 /// defines each byte.
1393 SmallVector<Value*, 8> ByteValues;
1394 ByteValues.resize(ITy->getBitWidth()/8);
1396 // Try to find all the pieces corresponding to the bswap.
1397 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
1398 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
1401 // Check to see if all of the bytes come from the same value.
1402 Value *V = ByteValues[0];
1403 if (V == 0) return 0; // Didn't find a byte? Must be zero.
1405 // Check to make sure that all of the bytes come from the same value.
1406 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
1407 if (ByteValues[i] != V)
1409 Module *M = I.getParent()->getParent()->getParent();
1410 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, ITy);
1411 return CallInst::Create(F, V);
1414 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
1415 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
1416 /// we can simplify this expression to "cond ? C : D or B".
1417 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
1418 Value *C, Value *D) {
1419 // If A is not a select of -1/0, this cannot match.
1421 if (!match(A, m_SExt(m_Value(Cond))) ||
1422 !Cond->getType()->isIntegerTy(1))
1425 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
1426 if (match(D, m_Not(m_SExt(m_Specific(Cond)))))
1427 return SelectInst::Create(Cond, C, B);
1428 if (match(D, m_SExt(m_Not(m_Specific(Cond)))))
1429 return SelectInst::Create(Cond, C, B);
1431 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
1432 if (match(B, m_Not(m_SExt(m_Specific(Cond)))))
1433 return SelectInst::Create(Cond, C, D);
1434 if (match(B, m_SExt(m_Not(m_Specific(Cond)))))
1435 return SelectInst::Create(Cond, C, D);
1439 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
1440 Value *InstCombiner::FoldOrOfICmps(ICmpInst *LHS, ICmpInst *RHS) {
1441 ICmpInst::Predicate LHSCC = LHS->getPredicate(), RHSCC = RHS->getPredicate();
1443 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
1444 if (PredicatesFoldable(LHSCC, RHSCC)) {
1445 if (LHS->getOperand(0) == RHS->getOperand(1) &&
1446 LHS->getOperand(1) == RHS->getOperand(0))
1447 LHS->swapOperands();
1448 if (LHS->getOperand(0) == RHS->getOperand(0) &&
1449 LHS->getOperand(1) == RHS->getOperand(1)) {
1450 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
1451 unsigned Code = getICmpCode(LHS) | getICmpCode(RHS);
1452 bool isSigned = LHS->isSigned() || RHS->isSigned();
1453 return getNewICmpValue(isSigned, Code, Op0, Op1, Builder);
1457 // handle (roughly):
1458 // (icmp ne (A & B), C) | (icmp ne (A & D), E)
1459 if (Value *V = foldLogOpOfMaskedICmps(LHS, RHS, ICmpInst::ICMP_NE, Builder))
1462 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
1463 Value *Val = LHS->getOperand(0), *Val2 = RHS->getOperand(0);
1464 ConstantInt *LHSCst = dyn_cast<ConstantInt>(LHS->getOperand(1));
1465 ConstantInt *RHSCst = dyn_cast<ConstantInt>(RHS->getOperand(1));
1466 if (LHSCst == 0 || RHSCst == 0) return 0;
1468 if (LHSCst == RHSCst && LHSCC == RHSCC) {
1469 // (icmp ne A, 0) | (icmp ne B, 0) --> (icmp ne (A|B), 0)
1470 if (LHSCC == ICmpInst::ICMP_NE && LHSCst->isZero()) {
1471 Value *NewOr = Builder->CreateOr(Val, Val2);
1472 return Builder->CreateICmp(LHSCC, NewOr, LHSCst);
1476 // (icmp ult (X + CA), C1) | (icmp eq X, C2) -> (icmp ule (X + CA), C1)
1477 // iff C2 + CA == C1.
1478 if (LHSCC == ICmpInst::ICMP_ULT && RHSCC == ICmpInst::ICMP_EQ) {
1479 ConstantInt *AddCst;
1480 if (match(Val, m_Add(m_Specific(Val2), m_ConstantInt(AddCst))))
1481 if (RHSCst->getValue() + AddCst->getValue() == LHSCst->getValue())
1482 return Builder->CreateICmpULE(Val, LHSCst);
1485 // From here on, we only handle:
1486 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
1487 if (Val != Val2) return 0;
1489 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
1490 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
1491 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
1492 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
1493 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
1496 // We can't fold (ugt x, C) | (sgt x, C2).
1497 if (!PredicatesFoldable(LHSCC, RHSCC))
1500 // Ensure that the larger constant is on the RHS.
1502 if (CmpInst::isSigned(LHSCC) ||
1503 (ICmpInst::isEquality(LHSCC) &&
1504 CmpInst::isSigned(RHSCC)))
1505 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
1507 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
1510 std::swap(LHS, RHS);
1511 std::swap(LHSCst, RHSCst);
1512 std::swap(LHSCC, RHSCC);
1515 // At this point, we know we have two icmp instructions
1516 // comparing a value against two constants and or'ing the result
1517 // together. Because of the above check, we know that we only have
1518 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
1519 // icmp folding check above), that the two constants are not
1521 assert(LHSCst != RHSCst && "Compares not folded above?");
1524 default: llvm_unreachable("Unknown integer condition code!");
1525 case ICmpInst::ICMP_EQ:
1527 default: llvm_unreachable("Unknown integer condition code!");
1528 case ICmpInst::ICMP_EQ:
1529 if (LHSCst == SubOne(RHSCst)) {
1530 // (X == 13 | X == 14) -> X-13 <u 2
1531 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
1532 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
1533 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
1534 return Builder->CreateICmpULT(Add, AddCST);
1536 break; // (X == 13 | X == 15) -> no change
1537 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
1538 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
1540 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
1541 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
1542 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
1546 case ICmpInst::ICMP_NE:
1548 default: llvm_unreachable("Unknown integer condition code!");
1549 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
1550 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
1551 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
1553 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
1554 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
1555 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
1556 return ConstantInt::getTrue(LHS->getContext());
1559 case ICmpInst::ICMP_ULT:
1561 default: llvm_unreachable("Unknown integer condition code!");
1562 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
1564 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
1565 // If RHSCst is [us]MAXINT, it is always false. Not handling
1566 // this can cause overflow.
1567 if (RHSCst->isMaxValue(false))
1569 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), false, false);
1570 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
1572 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
1573 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
1575 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
1579 case ICmpInst::ICMP_SLT:
1581 default: llvm_unreachable("Unknown integer condition code!");
1582 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
1584 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
1585 // If RHSCst is [us]MAXINT, it is always false. Not handling
1586 // this can cause overflow.
1587 if (RHSCst->isMaxValue(true))
1589 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst), true, false);
1590 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
1592 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
1593 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
1595 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
1599 case ICmpInst::ICMP_UGT:
1601 default: llvm_unreachable("Unknown integer condition code!");
1602 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
1603 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
1605 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
1607 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
1608 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
1609 return ConstantInt::getTrue(LHS->getContext());
1610 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
1614 case ICmpInst::ICMP_SGT:
1616 default: llvm_unreachable("Unknown integer condition code!");
1617 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
1618 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
1620 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
1622 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
1623 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
1624 return ConstantInt::getTrue(LHS->getContext());
1625 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
1633 /// FoldOrOfFCmps - Optimize (fcmp)|(fcmp). NOTE: Unlike the rest of
1634 /// instcombine, this returns a Value which should already be inserted into the
1636 Value *InstCombiner::FoldOrOfFCmps(FCmpInst *LHS, FCmpInst *RHS) {
1637 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
1638 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
1639 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
1640 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
1641 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
1642 // If either of the constants are nans, then the whole thing returns
1644 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
1645 return ConstantInt::getTrue(LHS->getContext());
1647 // Otherwise, no need to compare the two constants, compare the
1649 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1652 // Handle vector zeros. This occurs because the canonical form of
1653 // "fcmp uno x,x" is "fcmp uno x, 0".
1654 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
1655 isa<ConstantAggregateZero>(RHS->getOperand(1)))
1656 return Builder->CreateFCmpUNO(LHS->getOperand(0), RHS->getOperand(0));
1661 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
1662 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
1663 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
1665 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
1666 // Swap RHS operands to match LHS.
1667 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
1668 std::swap(Op1LHS, Op1RHS);
1670 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
1671 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
1673 return Builder->CreateFCmp((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
1674 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
1675 return ConstantInt::get(CmpInst::makeCmpResultType(LHS->getType()), 1);
1676 if (Op0CC == FCmpInst::FCMP_FALSE)
1678 if (Op1CC == FCmpInst::FCMP_FALSE)
1682 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
1683 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
1684 if (Op0Ordered == Op1Ordered) {
1685 // If both are ordered or unordered, return a new fcmp with
1686 // or'ed predicates.
1687 return getFCmpValue(Op0Ordered, Op0Pred|Op1Pred, Op0LHS, Op0RHS, Builder);
1693 /// FoldOrWithConstants - This helper function folds:
1695 /// ((A | B) & C1) | (B & C2)
1701 /// when the XOR of the two constants is "all ones" (-1).
1702 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
1703 Value *A, Value *B, Value *C) {
1704 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
1708 ConstantInt *CI2 = 0;
1709 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
1711 APInt Xor = CI1->getValue() ^ CI2->getValue();
1712 if (!Xor.isAllOnesValue()) return 0;
1714 if (V1 == A || V1 == B) {
1715 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
1716 return BinaryOperator::CreateOr(NewOp, V1);
1722 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
1723 bool Changed = SimplifyAssociativeOrCommutative(I);
1724 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1726 if (Value *V = SimplifyOrInst(Op0, Op1, TD))
1727 return ReplaceInstUsesWith(I, V);
1729 // (A&B)|(A&C) -> A&(B|C) etc
1730 if (Value *V = SimplifyUsingDistributiveLaws(I))
1731 return ReplaceInstUsesWith(I, V);
1733 // See if we can simplify any instructions used by the instruction whose sole
1734 // purpose is to compute bits we don't care about.
1735 if (SimplifyDemandedInstructionBits(I))
1738 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
1739 ConstantInt *C1 = 0; Value *X = 0;
1740 // (X & C1) | C2 --> (X | C2) & (C1|C2)
1741 // iff (C1 & C2) == 0.
1742 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
1743 (RHS->getValue() & C1->getValue()) != 0 &&
1745 Value *Or = Builder->CreateOr(X, RHS);
1747 return BinaryOperator::CreateAnd(Or,
1748 ConstantInt::get(I.getContext(),
1749 RHS->getValue() | C1->getValue()));
1752 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
1753 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
1755 Value *Or = Builder->CreateOr(X, RHS);
1757 return BinaryOperator::CreateXor(Or,
1758 ConstantInt::get(I.getContext(),
1759 C1->getValue() & ~RHS->getValue()));
1762 // Try to fold constant and into select arguments.
1763 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
1764 if (Instruction *R = FoldOpIntoSelect(I, SI))
1767 if (isa<PHINode>(Op0))
1768 if (Instruction *NV = FoldOpIntoPhi(I))
1772 Value *A = 0, *B = 0;
1773 ConstantInt *C1 = 0, *C2 = 0;
1775 // (A | B) | C and A | (B | C) -> bswap if possible.
1776 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
1777 if (match(Op0, m_Or(m_Value(), m_Value())) ||
1778 match(Op1, m_Or(m_Value(), m_Value())) ||
1779 (match(Op0, m_LogicalShift(m_Value(), m_Value())) &&
1780 match(Op1, m_LogicalShift(m_Value(), m_Value())))) {
1781 if (Instruction *BSwap = MatchBSwap(I))
1785 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
1786 if (Op0->hasOneUse() &&
1787 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1788 MaskedValueIsZero(Op1, C1->getValue())) {
1789 Value *NOr = Builder->CreateOr(A, Op1);
1791 return BinaryOperator::CreateXor(NOr, C1);
1794 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
1795 if (Op1->hasOneUse() &&
1796 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
1797 MaskedValueIsZero(Op0, C1->getValue())) {
1798 Value *NOr = Builder->CreateOr(A, Op0);
1800 return BinaryOperator::CreateXor(NOr, C1);
1804 Value *C = 0, *D = 0;
1805 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
1806 match(Op1, m_And(m_Value(B), m_Value(D)))) {
1807 Value *V1 = 0, *V2 = 0;
1808 C1 = dyn_cast<ConstantInt>(C);
1809 C2 = dyn_cast<ConstantInt>(D);
1810 if (C1 && C2) { // (A & C1)|(B & C2)
1811 // If we have: ((V + N) & C1) | (V & C2)
1812 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
1813 // replace with V+N.
1814 if (C1->getValue() == ~C2->getValue()) {
1815 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
1816 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
1817 // Add commutes, try both ways.
1818 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
1819 return ReplaceInstUsesWith(I, A);
1820 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
1821 return ReplaceInstUsesWith(I, A);
1823 // Or commutes, try both ways.
1824 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
1825 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
1826 // Add commutes, try both ways.
1827 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
1828 return ReplaceInstUsesWith(I, B);
1829 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
1830 return ReplaceInstUsesWith(I, B);
1834 if ((C1->getValue() & C2->getValue()) == 0) {
1835 // ((V | N) & C1) | (V & C2) --> (V|N) & (C1|C2)
1836 // iff (C1&C2) == 0 and (N&~C1) == 0
1837 if (match(A, m_Or(m_Value(V1), m_Value(V2))) &&
1838 ((V1 == B && MaskedValueIsZero(V2, ~C1->getValue())) || // (V|N)
1839 (V2 == B && MaskedValueIsZero(V1, ~C1->getValue())))) // (N|V)
1840 return BinaryOperator::CreateAnd(A,
1841 ConstantInt::get(A->getContext(),
1842 C1->getValue()|C2->getValue()));
1843 // Or commutes, try both ways.
1844 if (match(B, m_Or(m_Value(V1), m_Value(V2))) &&
1845 ((V1 == A && MaskedValueIsZero(V2, ~C2->getValue())) || // (V|N)
1846 (V2 == A && MaskedValueIsZero(V1, ~C2->getValue())))) // (N|V)
1847 return BinaryOperator::CreateAnd(B,
1848 ConstantInt::get(B->getContext(),
1849 C1->getValue()|C2->getValue()));
1851 // ((V|C3)&C1) | ((V|C4)&C2) --> (V|C3|C4)&(C1|C2)
1852 // iff (C1&C2) == 0 and (C3&~C1) == 0 and (C4&~C2) == 0.
1853 ConstantInt *C3 = 0, *C4 = 0;
1854 if (match(A, m_Or(m_Value(V1), m_ConstantInt(C3))) &&
1855 (C3->getValue() & ~C1->getValue()) == 0 &&
1856 match(B, m_Or(m_Specific(V1), m_ConstantInt(C4))) &&
1857 (C4->getValue() & ~C2->getValue()) == 0) {
1858 V2 = Builder->CreateOr(V1, ConstantExpr::getOr(C3, C4), "bitfield");
1859 return BinaryOperator::CreateAnd(V2,
1860 ConstantInt::get(B->getContext(),
1861 C1->getValue()|C2->getValue()));
1866 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants.
1867 // Don't do this for vector select idioms, the code generator doesn't handle
1869 if (!I.getType()->isVectorTy()) {
1870 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D))
1872 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C))
1874 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D))
1876 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C))
1880 // ((A&~B)|(~A&B)) -> A^B
1881 if ((match(C, m_Not(m_Specific(D))) &&
1882 match(B, m_Not(m_Specific(A)))))
1883 return BinaryOperator::CreateXor(A, D);
1884 // ((~B&A)|(~A&B)) -> A^B
1885 if ((match(A, m_Not(m_Specific(D))) &&
1886 match(B, m_Not(m_Specific(C)))))
1887 return BinaryOperator::CreateXor(C, D);
1888 // ((A&~B)|(B&~A)) -> A^B
1889 if ((match(C, m_Not(m_Specific(B))) &&
1890 match(D, m_Not(m_Specific(A)))))
1891 return BinaryOperator::CreateXor(A, B);
1892 // ((~B&A)|(B&~A)) -> A^B
1893 if ((match(A, m_Not(m_Specific(B))) &&
1894 match(D, m_Not(m_Specific(C)))))
1895 return BinaryOperator::CreateXor(C, B);
1897 // ((A|B)&1)|(B&-2) -> (A&1) | B
1898 if (match(A, m_Or(m_Value(V1), m_Specific(B))) ||
1899 match(A, m_Or(m_Specific(B), m_Value(V1)))) {
1900 Instruction *Ret = FoldOrWithConstants(I, Op1, V1, B, C);
1901 if (Ret) return Ret;
1903 // (B&-2)|((A|B)&1) -> (A&1) | B
1904 if (match(B, m_Or(m_Specific(A), m_Value(V1))) ||
1905 match(B, m_Or(m_Value(V1), m_Specific(A)))) {
1906 Instruction *Ret = FoldOrWithConstants(I, Op0, A, V1, D);
1907 if (Ret) return Ret;
1911 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
1912 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
1913 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
1914 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
1915 SI0->getOperand(1) == SI1->getOperand(1) &&
1916 (SI0->hasOneUse() || SI1->hasOneUse())) {
1917 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
1919 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
1920 SI1->getOperand(1));
1924 // (~A | ~B) == (~(A & B)) - De Morgan's Law
1925 if (Value *Op0NotVal = dyn_castNotVal(Op0))
1926 if (Value *Op1NotVal = dyn_castNotVal(Op1))
1927 if (Op0->hasOneUse() && Op1->hasOneUse()) {
1928 Value *And = Builder->CreateAnd(Op0NotVal, Op1NotVal,
1929 I.getName()+".demorgan");
1930 return BinaryOperator::CreateNot(And);
1933 // Canonicalize xor to the RHS.
1934 if (match(Op0, m_Xor(m_Value(), m_Value())))
1935 std::swap(Op0, Op1);
1937 // A | ( A ^ B) -> A | B
1938 // A | (~A ^ B) -> A | ~B
1939 if (match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
1940 if (Op0 == A || Op0 == B)
1941 return BinaryOperator::CreateOr(A, B);
1943 if (Op1->hasOneUse() && match(A, m_Not(m_Specific(Op0)))) {
1944 Value *Not = Builder->CreateNot(B, B->getName()+".not");
1945 return BinaryOperator::CreateOr(Not, Op0);
1947 if (Op1->hasOneUse() && match(B, m_Not(m_Specific(Op0)))) {
1948 Value *Not = Builder->CreateNot(A, A->getName()+".not");
1949 return BinaryOperator::CreateOr(Not, Op0);
1953 // A | ~(A | B) -> A | ~B
1954 // A | ~(A ^ B) -> A | ~B
1955 if (match(Op1, m_Not(m_Value(A))))
1956 if (BinaryOperator *B = dyn_cast<BinaryOperator>(A))
1957 if ((Op0 == B->getOperand(0) || Op0 == B->getOperand(1)) &&
1958 Op1->hasOneUse() && (B->getOpcode() == Instruction::Or ||
1959 B->getOpcode() == Instruction::Xor)) {
1960 Value *NotOp = Op0 == B->getOperand(0) ? B->getOperand(1) :
1962 Value *Not = Builder->CreateNot(NotOp, NotOp->getName()+".not");
1963 return BinaryOperator::CreateOr(Not, Op0);
1966 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
1967 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
1968 if (Value *Res = FoldOrOfICmps(LHS, RHS))
1969 return ReplaceInstUsesWith(I, Res);
1971 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
1972 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0)))
1973 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
1974 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
1975 return ReplaceInstUsesWith(I, Res);
1977 // fold (or (cast A), (cast B)) -> (cast (or A, B))
1978 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
1979 CastInst *Op1C = dyn_cast<CastInst>(Op1);
1980 if (Op1C && Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
1981 Type *SrcTy = Op0C->getOperand(0)->getType();
1982 if (SrcTy == Op1C->getOperand(0)->getType() &&
1983 SrcTy->isIntOrIntVectorTy()) {
1984 Value *Op0COp = Op0C->getOperand(0), *Op1COp = Op1C->getOperand(0);
1986 if ((!isa<ICmpInst>(Op0COp) || !isa<ICmpInst>(Op1COp)) &&
1987 // Only do this if the casts both really cause code to be
1989 ShouldOptimizeCast(Op0C->getOpcode(), Op0COp, I.getType()) &&
1990 ShouldOptimizeCast(Op1C->getOpcode(), Op1COp, I.getType())) {
1991 Value *NewOp = Builder->CreateOr(Op0COp, Op1COp, I.getName());
1992 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
1995 // If this is or(cast(icmp), cast(icmp)), try to fold this even if the
1996 // cast is otherwise not optimizable. This happens for vector sexts.
1997 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1COp))
1998 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0COp))
1999 if (Value *Res = FoldOrOfICmps(LHS, RHS))
2000 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2002 // If this is or(cast(fcmp), cast(fcmp)), try to fold this even if the
2003 // cast is otherwise not optimizable. This happens for vector sexts.
2004 if (FCmpInst *RHS = dyn_cast<FCmpInst>(Op1COp))
2005 if (FCmpInst *LHS = dyn_cast<FCmpInst>(Op0COp))
2006 if (Value *Res = FoldOrOfFCmps(LHS, RHS))
2007 return CastInst::Create(Op0C->getOpcode(), Res, I.getType());
2012 // or(sext(A), B) -> A ? -1 : B where A is an i1
2013 // or(A, sext(B)) -> B ? -1 : A where B is an i1
2014 if (match(Op0, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2015 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op1);
2016 if (match(Op1, m_SExt(m_Value(A))) && A->getType()->isIntegerTy(1))
2017 return SelectInst::Create(A, ConstantInt::getSigned(I.getType(), -1), Op0);
2019 // Note: If we've gotten to the point of visiting the outer OR, then the
2020 // inner one couldn't be simplified. If it was a constant, then it won't
2021 // be simplified by a later pass either, so we try swapping the inner/outer
2022 // ORs in the hopes that we'll be able to simplify it this way.
2023 // (X|C) | V --> (X|V) | C
2024 if (Op0->hasOneUse() && !isa<ConstantInt>(Op1) &&
2025 match(Op0, m_Or(m_Value(A), m_ConstantInt(C1)))) {
2026 Value *Inner = Builder->CreateOr(A, Op1);
2027 Inner->takeName(Op0);
2028 return BinaryOperator::CreateOr(Inner, C1);
2031 return Changed ? &I : 0;
2034 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
2035 bool Changed = SimplifyAssociativeOrCommutative(I);
2036 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2038 if (Value *V = SimplifyXorInst(Op0, Op1, TD))
2039 return ReplaceInstUsesWith(I, V);
2041 // (A&B)^(A&C) -> A&(B^C) etc
2042 if (Value *V = SimplifyUsingDistributiveLaws(I))
2043 return ReplaceInstUsesWith(I, V);
2045 // See if we can simplify any instructions used by the instruction whose sole
2046 // purpose is to compute bits we don't care about.
2047 if (SimplifyDemandedInstructionBits(I))
2050 // Is this a ~ operation?
2051 if (Value *NotOp = dyn_castNotVal(&I)) {
2052 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
2053 if (Op0I->getOpcode() == Instruction::And ||
2054 Op0I->getOpcode() == Instruction::Or) {
2055 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
2056 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
2057 if (dyn_castNotVal(Op0I->getOperand(1)))
2058 Op0I->swapOperands();
2059 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
2061 Builder->CreateNot(Op0I->getOperand(1),
2062 Op0I->getOperand(1)->getName()+".not");
2063 if (Op0I->getOpcode() == Instruction::And)
2064 return BinaryOperator::CreateOr(Op0NotVal, NotY);
2065 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
2068 // ~(X & Y) --> (~X | ~Y) - De Morgan's Law
2069 // ~(X | Y) === (~X & ~Y) - De Morgan's Law
2070 if (isFreeToInvert(Op0I->getOperand(0)) &&
2071 isFreeToInvert(Op0I->getOperand(1))) {
2073 Builder->CreateNot(Op0I->getOperand(0), "notlhs");
2075 Builder->CreateNot(Op0I->getOperand(1), "notrhs");
2076 if (Op0I->getOpcode() == Instruction::And)
2077 return BinaryOperator::CreateOr(NotX, NotY);
2078 return BinaryOperator::CreateAnd(NotX, NotY);
2081 } else if (Op0I->getOpcode() == Instruction::AShr) {
2082 // ~(~X >>s Y) --> (X >>s Y)
2083 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0)))
2084 return BinaryOperator::CreateAShr(Op0NotVal, Op0I->getOperand(1));
2090 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2091 if (RHS->isOne() && Op0->hasOneUse())
2092 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
2093 if (CmpInst *CI = dyn_cast<CmpInst>(Op0))
2094 return CmpInst::Create(CI->getOpcode(),
2095 CI->getInversePredicate(),
2096 CI->getOperand(0), CI->getOperand(1));
2098 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
2099 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2100 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
2101 if (CI->hasOneUse() && Op0C->hasOneUse()) {
2102 Instruction::CastOps Opcode = Op0C->getOpcode();
2103 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
2104 (RHS == ConstantExpr::getCast(Opcode,
2105 ConstantInt::getTrue(I.getContext()),
2106 Op0C->getDestTy()))) {
2107 CI->setPredicate(CI->getInversePredicate());
2108 return CastInst::Create(Opcode, CI, Op0C->getType());
2114 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2115 // ~(c-X) == X-c-1 == X+(-c-1)
2116 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
2117 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
2118 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
2119 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
2120 ConstantInt::get(I.getType(), 1));
2121 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
2124 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
2125 if (Op0I->getOpcode() == Instruction::Add) {
2126 // ~(X-c) --> (-c-1)-X
2127 if (RHS->isAllOnesValue()) {
2128 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
2129 return BinaryOperator::CreateSub(
2130 ConstantExpr::getSub(NegOp0CI,
2131 ConstantInt::get(I.getType(), 1)),
2132 Op0I->getOperand(0));
2133 } else if (RHS->getValue().isSignBit()) {
2134 // (X + C) ^ signbit -> (X + C + signbit)
2135 Constant *C = ConstantInt::get(I.getContext(),
2136 RHS->getValue() + Op0CI->getValue());
2137 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
2140 } else if (Op0I->getOpcode() == Instruction::Or) {
2141 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
2142 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
2143 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
2144 // Anything in both C1 and C2 is known to be zero, remove it from
2146 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
2147 NewRHS = ConstantExpr::getAnd(NewRHS,
2148 ConstantExpr::getNot(CommonBits));
2150 I.setOperand(0, Op0I->getOperand(0));
2151 I.setOperand(1, NewRHS);
2158 // Try to fold constant and into select arguments.
2159 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2160 if (Instruction *R = FoldOpIntoSelect(I, SI))
2162 if (isa<PHINode>(Op0))
2163 if (Instruction *NV = FoldOpIntoPhi(I))
2167 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
2170 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
2171 if (A == Op0) { // B^(B|A) == (A|B)^B
2172 Op1I->swapOperands();
2174 std::swap(Op0, Op1);
2175 } else if (B == Op0) { // B^(A|B) == (A|B)^B
2176 I.swapOperands(); // Simplified below.
2177 std::swap(Op0, Op1);
2179 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
2181 if (A == Op0) { // A^(A&B) -> A^(B&A)
2182 Op1I->swapOperands();
2185 if (B == Op0) { // A^(B&A) -> (B&A)^A
2186 I.swapOperands(); // Simplified below.
2187 std::swap(Op0, Op1);
2192 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
2195 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2196 Op0I->hasOneUse()) {
2197 if (A == Op1) // (B|A)^B == (A|B)^B
2199 if (B == Op1) // (A|B)^B == A & ~B
2200 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1));
2201 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2203 if (A == Op1) // (A&B)^A -> (B&A)^A
2205 if (B == Op1 && // (B&A)^A == ~B & A
2206 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
2207 return BinaryOperator::CreateAnd(Builder->CreateNot(A), Op1);
2212 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
2213 if (Op0I && Op1I && Op0I->isShift() &&
2214 Op0I->getOpcode() == Op1I->getOpcode() &&
2215 Op0I->getOperand(1) == Op1I->getOperand(1) &&
2216 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
2218 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
2220 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
2221 Op1I->getOperand(1));
2225 Value *A, *B, *C, *D;
2226 // (A & B)^(A | B) -> A ^ B
2227 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
2228 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
2229 if ((A == C && B == D) || (A == D && B == C))
2230 return BinaryOperator::CreateXor(A, B);
2232 // (A | B)^(A & B) -> A ^ B
2233 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
2234 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
2235 if ((A == C && B == D) || (A == D && B == C))
2236 return BinaryOperator::CreateXor(A, B);
2240 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
2241 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
2242 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
2243 if (PredicatesFoldable(LHS->getPredicate(), RHS->getPredicate())) {
2244 if (LHS->getOperand(0) == RHS->getOperand(1) &&
2245 LHS->getOperand(1) == RHS->getOperand(0))
2246 LHS->swapOperands();
2247 if (LHS->getOperand(0) == RHS->getOperand(0) &&
2248 LHS->getOperand(1) == RHS->getOperand(1)) {
2249 Value *Op0 = LHS->getOperand(0), *Op1 = LHS->getOperand(1);
2250 unsigned Code = getICmpCode(LHS) ^ getICmpCode(RHS);
2251 bool isSigned = LHS->isSigned() || RHS->isSigned();
2252 return ReplaceInstUsesWith(I,
2253 getNewICmpValue(isSigned, Code, Op0, Op1,
2258 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
2259 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
2260 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
2261 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
2262 Type *SrcTy = Op0C->getOperand(0)->getType();
2263 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegerTy() &&
2264 // Only do this if the casts both really cause code to be generated.
2265 ShouldOptimizeCast(Op0C->getOpcode(), Op0C->getOperand(0),
2267 ShouldOptimizeCast(Op1C->getOpcode(), Op1C->getOperand(0),
2269 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
2270 Op1C->getOperand(0), I.getName());
2271 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
2276 return Changed ? &I : 0;